LunarGate – a self-hosted OpenAI-compatible LLM gateway
Go gateway with circuit breakers, but auth isn't production-ready yet.
Stop being locked into one LLM provider. UnifyRoute is a self-hosted gateway that routes, fails over, and manages quotas across OpenAI, Anthropic, and more with a drop-in OpenAI-compatible API.
Drop-in OpenAI API gateway with failover—LiteLLM does this but this has a dashboard.
Developers running production LLM-powered applications
LiteLLM · Helicone · Portkey
I built UnifyRoute because I kept running into the same problem: rate limits, quota exhaustion, and provider outages were breaking my LLM-powered apps at the worst times.
UnifyRoute is a self-hosted gateway that sits in front of your LLM providers (OpenAI, Anthropic, etc.) and handles routing, failover, and quota management automatically — with a fully OpenAI-compatible API, so you don't change a single line of your existing code.
What it does: - Drop-in OpenAI-compatible API (/chat/completions, /models, etc.) - Tier-based routing: define which providers to try and in what order - Automatic failover when a provider fails or hits quota - Web dashboard to manage providers, credentials, and usage - Self-hosted — your API keys never leave your infrastructure - Works with any tool that supports OpenAI's API (LangChain, LlamaIndex, etc.)
Quick start (Docker): git clone https://github.com/unifyroute/UnifyRoute.git cd UnifyRoute && cp sample.env .env ./unifyroute setup && ./unifyroute start # Dashboard at http://localhost:6565
It's open source under Apache 2.0.
Happy to answer questions about the architecture or design decisions.
Go gateway with circuit breakers, but auth isn't production-ready yet.
Free AI gateway with failover, but LiteLLM already aggregates providers with better maturity.
Zero-trust networking via zrok beats LiteLLM when your GPUs sit behind NAT.
Multi-backend LLM manager when Ollama and LM Studio already handle this.
Free-tier failover across Gemini, Groq, Mistral beats single-provider rate limits.
Prompt injection detection at 100% precision — but only catches 43% of actual injections.